What it is
Builds per-student LLM simulators from sparse learner data via pooled pretraining plus student-specific tuning, and evaluates them on fidelity and response to tutor guidance. Unlike prompted LLM role-play or state-tracking models that get only one axis right, it trains individualized simulators with pooled domain training then per-student specialization to match a student's behavior and update under tutor guidance.
Key results
- Chess: F=0.51, R=0.91 vs GPT-5.4's 0.23, 0.72
- Chess: beats Maia2 on both axes, 0.51/0.91 vs 0.45/0.27
- RL tutor using StudentSim reward was rated above no-RL and GPT-5.4-reward RL
Training on real learner records with shared pretraining plus per-student tuning can jointly recover student fidelity and teachability.